Skin Cancer Classification using Convolutional Neural Network and Feature Selection Strategy

Abstract

The American Cancer Society states that the most common illness among people is skin cancer. It is mainly diagnosed visually, starting with a clinical screening and maybe requiring a histological examination, biopsy, and dermoscopic analysis. Skin cancer is caused by mistakes (mutations) in the DNA of skin cells. In the realm of medical science, identifying a range of local and gene-related illness diseases depends on the ability to classify skin lesions into distinct malignant types. Despite the fact that skin cancers are the worst type of the disease; individuals who are detected with it early on frequently recover completely. Different writers have investigated a number of methods for automatic diagnosis and detection utilizing a variety of techniques. This thesis uses CNN and HOG feature selectors with transfer learning for dermoscopy pictures to demonstrate the classification of skin lesions in those cases where skin cancer is mentioned. This research uses the Kaggle 2018 challenge dataset. This study attempted to use convolutional neural networks to classify photos of skin lesions. Deep neural networks demonstrate enormous promise for classifying images while accounting for the significant environmental heterogeneity. The mutations cause the cells to grow out of control, resulting in a mass of malignant cells. The division of these lesions into various malignant types, including benign keratosis (BKL), actinic keratosis (AKIEC), basal cell carcinoma (BCC), melanoma (MEL), melanomic neves (NV), and vascular lesions (VASC), offers some insight into the problem. The process utilized in this work produces the best results when we mix the CNN model, HOG feature selector, and transfer learning. The accuracy we obtain is 96.0%.